deadlock-prevention（预防死锁）

Deadlock Prevention

Overview

When two processes wait for each other forever, you have a deadlock.

Deadlocks are permanent freezes—no crash, no error, just silence. They occur when processes hold resources while waiting for resources held by others, creating circular wait.

When to Use

• Acquiring multiple locks or resources
• Nested transactions across tables/services
• Distributed systems with cross-service calls
• "System freezes under load"
• No errors but operations never complete

The Iron Rule

NEVER acquire locks in inconsistent order across the codebase.

No exceptions:

• Not for "this path is rare"
• Not for "we'll add timeout"
• Not for "different services, won't conflict"
• Not for "only happens under heavy load"

Inconsistent lock ordering guarantees eventual deadlock.

The Four Deadlock Conditions

Deadlock requires ALL four. Break ANY one to prevent:

┌────────────────────────────────────────────────────────────┐
│ 1. MUTUAL EXCLUSION    │ Resource can only be held by one │
│ 2. HOLD AND WAIT       │ Hold one, wait for another       │
│ 3. NO PREEMPTION       │ Can't force release of resource  │
│ 4. CIRCULAR WAIT       │ A waits for B, B waits for A     │
└────────────────────────────────────────────────────────────┘

Most practical: Break Circular Wait with consistent ordering.

Detection: The Circular Wait Pattern

// ❌ VIOLATION: Inconsistent lock order = guaranteed deadlock

// Service A: Transfer from X to Y
async function transferAtoB(fromId: string, toId: string, amount: number) {
  await lockAccount(fromId);  // Lock A first
  await lockAccount(toId);    // Then lock B
  // transfer...
  await unlockAccount(toId);
  await unlockAccount(fromId);
}

// Service B: Transfer from Y to X
async function transferBtoA(fromId: string, toId: string, amount: number) {
  await lockAccount(fromId);  // Lock B first (OPPOSITE ORDER!)
  await lockAccount(toId);    // Then lock A
  // transfer...
}

// Deadlock scenario:
// T1: Locks account A, waits for B
// T2: Locks account B, waits for A
// Result: Both wait forever

Prevention Strategies

1. Consistent Lock Ordering (Primary Strategy)

// ✅ CORRECT: Always acquire locks in the same order
async function transfer(
  accountId1: string,
  accountId2: string,
  amount: number
): Promise<void> {
  // ALWAYS lock lower ID first - creates total ordering
  const [first, second] = accountId1 < accountId2
    ? [accountId1, accountId2]
    : [accountId2, accountId1];

  const lock1 = await acquireLock(`account:${first}`);
  const lock2 = await acquireLock(`account:${second}`);

  try {
    await doTransfer(accountId1, accountId2, amount);
  } finally {
    await releaseLock(lock2);
    await releaseLock(lock1);  // Release in reverse order
  }
}

2. Lock Timeout (Break No Preemption)

// ✅ CORRECT: Timeout prevents permanent deadlock
async function acquireWithTimeout(
  resource: string,
  timeoutMs: number
): Promise<Lock | null> {
  const deadline = Date.now() + timeoutMs;

  while (Date.now() < deadline) {
    const lock = await tryAcquireLock(resource);
    if (lock) return lock;

    // Backoff with jitter
    await sleep(50 + Math.random() * 50);
  }

  return null;  // Timeout - caller must handle
}

async function transfer(from: string, to: string, amount: number) {
  const lock1 = await acquireWithTimeout(`account:${from}`, 5000);
  if (!lock1) throw new LockTimeoutError(from);

  try {
    const lock2 = await acquireWithTimeout(`account:${to}`, 5000);
    if (!lock2) {
      throw new LockTimeoutError(to);  // Release lock1 in finally
    }

    try {
      await doTransfer(from, to, amount);
    } finally {
      await releaseLock(lock2);
    }
  } finally {
    await releaseLock(lock1);
  }
}

3. Try-Lock with Backoff (Break Hold and Wait)

// ✅ CORRECT: Don't hold while waiting - release and retry
async function acquireMultipleLocks(
  resources: string[],
  maxRetries = 10
): Promise<Lock[]> {
  for (let attempt = 0; attempt < maxRetries; attempt++) {
    const locks: Lock[] = [];
    let success = true;

    // Sort for consistent ordering
    const sorted = [...resources].sort();

    for (const resource of sorted) {
      const lock = await tryAcquireLock(resource);
      if (lock) {
        locks.push(lock);
      } else {
        // Can't get this lock - release all and retry
        success = false;
        break;
      }
    }

    if (success) {
      return locks;
    }

    // Release all acquired locks
    await Promise.all(locks.map(releaseLock));

    // Exponential backoff with jitter
    await sleep(Math.pow(2, attempt) * 10 + Math.random() * 50);
  }

  throw new DeadlockPreventionError('Could not acquire all locks');
}

4. Single Lock for Related Resources

// ✅ CORRECT: One lock covers all related resources
async function transfer(from: string, to: string, amount: number) {
  // Use a higher-level lock instead of per-account locks
  const transferLock = await acquireLock('transfer-system');

  try {
    // No deadlock possible - only one lock
    await doTransfer(from, to, amount);
  } finally {
    await releaseLock(transferLock);
  }
}

// Trade-off: Less parallelism, but no deadlock risk
// Use when: Operations are fast, correctness > throughput

5. Lock Hierarchy

// ✅ CORRECT: Hierarchical locking - always acquire parent before child
enum LockLevel {
  SYSTEM = 0,
  REGION = 1,
  ACCOUNT = 2,
  TRANSACTION = 3,
}

class HierarchicalLock {
  private heldLocks: Map<LockLevel, Set<string>> = new Map();

  async acquire(level: LockLevel, resource: string): Promise<void> {
    // Validate hierarchy - can only acquire lower levels
    for (const [heldLevel] of this.heldLocks) {
      if (level <= heldLevel) {
        throw new LockHierarchyViolation(
          `Cannot acquire ${level} while holding ${heldLevel}`
        );
      }
    }

    await this.doAcquire(level, resource);

    if (!this.heldLocks.has(level)) {
      this.heldLocks.set(level, new Set());
    }
    this.heldLocks.get(level)!.add(resource);
  }
}

Database-Specific Patterns

SQL: Lock Ordering

-- ❌ VIOLATION: No consistent order
-- Transaction 1
SELECT * FROM accounts WHERE id = 'A' FOR UPDATE;
SELECT * FROM accounts WHERE id = 'B' FOR UPDATE;

-- Transaction 2 (opposite order)
SELECT * FROM accounts WHERE id = 'B' FOR UPDATE;
SELECT * FROM accounts WHERE id = 'A' FOR UPDATE;

-- ✅ CORRECT: Always order by ID
SELECT * FROM accounts
WHERE id IN ('A', 'B')
ORDER BY id
FOR UPDATE;

PostgreSQL: Lock Timeout

-- Set lock timeout to prevent permanent deadlock
SET lock_timeout = '5s';

BEGIN;
SELECT * FROM accounts WHERE id = 'A' FOR UPDATE;
-- If can't acquire within 5s, transaction fails with error
-- Better than hanging forever
COMMIT;

Distributed System Patterns

// ✅ CORRECT: Distributed lock with lease
class DistributedLock {
  async acquire(resource: string, leaseDuration: number): Promise<Lease> {
    const lease: Lease = {
      id: uuid(),
      resource,
      expiresAt: Date.now() + leaseDuration,
    };

    // Atomic set-if-not-exists with TTL
    const acquired = await redis.set(
      `lock:${resource}`,
      JSON.stringify(lease),
      'NX',  // Only if not exists
      'PX', leaseDuration  // Expire after duration
    );

    if (!acquired) {
      throw new LockNotAvailableError(resource);
    }

    return lease;
  }

  async release(lease: Lease): Promise<void> {
    // Only release if we still own it (compare lease ID)
    const script = `
      if redis.call("get", KEYS[1]) == ARGV[1] then
        return redis.call("del", KEYS[1])
      else
        return 0
      end
    `;
    await redis.eval(script, 1, `lock:${lease.resource}`, JSON.stringify(lease));
  }
}

Pressure Resistance Protocol

1. "Just Add a Longer Timeout"

Pressure: "Increase timeout to avoid failures"

Response: Longer timeout = longer deadlock before failure. The deadlock still exists.

Action: Fix the ordering. Timeout is safety net, not solution.

2. "It Only Happens Under Heavy Load"

Pressure: "Rare edge case, not worth fixing"

Response: Production IS heavy load. "Rare" becomes "constant" at scale. Deadlocks are permanent failures.

Action: Fix now. Deadlocks scale with traffic.

3. "Different Services Won't Conflict"

Pressure: "Services are independent"

Response: If they share resources (DB rows, Redis keys, files), they conflict.

Action: Coordinate lock ordering across services.

4. "We Can Just Restart"

Pressure: "Restart clears deadlocks"

Response: Restart kills in-flight transactions. Data corruption risk. User impact. Root cause remains.

Action: Prevent deadlocks. Don't treat symptoms.

Red Flags - STOP and Reconsider

If you notice ANY of these, deadlock risk is high:

• Acquiring locks in different orders across code paths
• Nested lock acquisition without clear hierarchy
• "This endpoint hangs sometimes"
• No lock timeouts
• Locks held across async operations
• Cross-service locking without coordination
• Locks acquired in loops

All of these mean: Redesign lock strategy.

Deadlock Detection

// Runtime deadlock detection
class DeadlockDetector {
  private waitGraph: Map<string, Set<string>> = new Map();  // who waits for whom

  recordWaiting(waiter: string, holder: string): void {
    if (!this.waitGraph.has(waiter)) {
      this.waitGraph.set(waiter, new Set());
    }
    this.waitGraph.get(waiter)!.add(holder);

    // Check for cycle
    if (this.hasCycle(waiter)) {
      throw new DeadlockDetectedError(
        `Deadlock: ${this.describeCycle(waiter)}`
      );
    }
  }

  private hasCycle(start: string, visited = new Set<string>()): boolean {
    if (visited.has(start)) return true;
    visited.add(start);

    for (const next of this.waitGraph.get(start) || []) {
      if (this.hasCycle(next, visited)) return true;
    }

    visited.delete(start);
    return false;
  }
}

Common Rationalizations (All Invalid)

Quick Reference

Deadlock Prevention Checklist

The Bottom Line

Order your locks. Always. Every path. Every service.

Deadlocks come from inconsistent lock ordering. Pick an order (alphabetical, ID-based, hierarchical) and enforce it everywhere. Add timeouts as safety nets. Monitor for lock contention. When pressured to "just add timeout," fix the ordering instead.